Test and Evaluation of Neural Network Applications for Seismic Signal Discrimination

Abstract : This study describes operational test and evaluation of two neural network applications that were integrated into the Intelligent Monitoring System (IMS) for automated processing and interpretation of regional seismic data. Also reported is the result of a preliminary study on the application of neural networks to regional seismic event identification. The first application is for initial identification of seismic phases (P or S) recorded by 3-component stations based on polarization data and context. This neural network performed 3-6% better than the current rule-based system when tested on data obtained from the 3-component IRIS stations in the former Soviet Union. This resulted in an improved event bulletin which showed that the number of analyst-verified events that were missed by the automated processing decreased by more than a factor of 2 (about 10 events/week). The second operational test was conducted on the neural network developed by MIT/Lincoln Laboratory for regional final phase identification (e.g., Pn, Pg, Sn, Lg, and Rg). This neural network performed 3. 3% better than the rule-based system in IMS station processing. However, for the final phase identifications obtained after network processing (where data from all stations are combined), the gain dropped to about 1.0%. It is likely that this could be regained by using the neural network phase identification confidence factors in the network processing. Finally, our preliminary study on the application of neural networks to identify regional seismic events on the basis of coda shape gave about 80% accuracy on data recorded at GERESS. In general, the neural network classifier utilized the coda decay rate which was lower for the earthquakes than it was for the explosions, although there was substantial overlap.